sh hour and confirmed 
mplex situations, The 
ion is consisted of two 
5), set about one meter 
> rate is set at 75 
stereo synchronization 
as taken from a point 
y (figure 5). 
  
Platform #1 
    
South Exit 
me initial values and 
of particles as N=500. 
| position of people 
z). The size of the 
[m] and d=0.3[m] 
the initial velocity of 
system model, we set 
r some trials. Finally 
id ground coordinate 
frames). During this 
al are to be tracked. 
proposed method by 
tained from tracking 
ge. As a result, we 
and 40 people of 44 
10). 
1 by system model 
  
Success £ of 
SS person 
tracked to the 
ticket gate 
40 / 44 
  
  
35/44 
28/44 
; in the image show 
king. The numbers 
Jue number given to 
> tracked person. 
  
  
  
  
  
  
Frame #35 
  
  
iud. A 5 Frame #40 
E 
    
  
Figure 7. Results 
We confirm that under the situation without occlusion and 
proximity between people or people and object, tracking 
succeeded in almost all frames. In addition, in situations 
involving a speed change, occlusion, proximity, avoidance 
behavior and direction change at around the gate, success rate 
stays at a high level. For example, two people shown as an oval 
of red and blue on figure 7 are successfully tracked under the 
condition that they are changing the direction and avoiding the 
collision near the ticket gate. 
Although the effectiveness of the proposed method has been 
shown from the results above, some points to be improved 
remain for more accurate tracking. For example, by introducing 
the interaction between person and object to system model, the 
accuracy when people pass through the ticket gate may improve. 
In addition, considering the interaction with the person beside 
and behind or minimum distance between people in system 
model would bring a more robust tracking. 
5.3 Comparison with Other System Model 
To verify the effectiveness of the proposed model, we use other 
System models and compare tracking results. Two cases are 
experimented: (a) system model with noise term only (v,.,-0 in 
equation (5)) and (b) system model with destination term (use 
term (b) in equation (4) and destinations are set manually). In 
case of (a), success rate dropped to 53% and the number of 
people tracked to the ticket gate to 35. This shows that 
integration of pedestrian behavior model with tracking method 
Is meaningful. In the same way, in case of (b), the result is 66% 
and 28 people. Failure cases are mainly caused by direction 
change at the ticket gate, for the direction choice around there is 
not necessarily the same as the final destination (table 6). 
5.4 Acquisition of Passenger Flow 
We can get passenger flow information by projecting the 
tracking result to the ground floor. Figure 8 shows a part of the 
acquired passenger flow. From this flow information, we try to 
get passenger’s ticket gate choice automatically. This data are 
more useful than simple cross-sectional data because each 
passenger’s origin is related to the choice of the ticket gate. The 
result is shown in table 9. Compared with the data acquired 
manually, 37 of 42 people’s choices are successfully obtained. 
Another example shown in table 10 is OD data. 26 of 39 
person's OD data are correctly acquired. From this result we can 
grasp the general tendency like the flow between south exit and 
platform 2 is at a high level. In this way, the proposed method 
increases the possibility to acquire detail flow data of the 
individuals. It is expected that comprehension of people’s 
behavior using this flow data leads to more sophisticated and 
precise flow control and facility design. 
5.5 Integration with Detection Method 
We try to expand the method to achieve the long time tracking, 
integrating a detection method of people entering the image, 
instead of setting it manually. After person is detected at time f, 
we assume a probability distribution p(x,) and forecast the state 
X,+1 by system model with noise term only. Then we filter x;